Deep learning (DL) is revolutionizing data analytics applications across many domains. But making effective use of DL is often a painful empirical process, since accuracy is tied to the data representation, neural architecture, and hyper-parameter settings. This process, called model selection, is a bottleneck to democratizing DL due to both resource costs and user time spent.

Cerebro is a first-of-its-kind platform that mitigates this bottleneck for DL model selection at scale. It raises model building throughput without raising resource costs and while ensuring accuracy, reproducibility, and generality to support multiple DL frameworks (PyTorch and TensorFlow). Our target setting is small clusters, which covers the vast majority of DL use cases in practice.